Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Neur2BiLO: Neural Bilevel Optimization
Authors: Justin Dumouchelle, Esther Julien, Jannis Kurtz, Elias Khalil
NeurIPS 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through a series of experiments on (i) the bilevel knapsack interdiction problem, (ii) the critical node problem from network security, (iii) a donor-recipient healthcare problem, and (iv) the DNDP, we will show that NEUR2BILO is easy to train and produces, very quickly, heuristic solutions that are competitive with state-of-the-art methods. |
| Researcher Affiliation | Academia | Justin Dumouchelle University of Toronto Esther Julien TU Delft Jannis Kurtz University of Amsterdam Elias B. Khalil University of Toronto |
| Pseudocode | Yes | Algorithm 1 NEUR2BILO Data Collection and Training" and "Algorithm 2 NEUR2BILO Optimization" are present in Appendix B. |
| Open Source Code | Yes | Our code and data are available at https://github.com/khalil-research/Neur2Bi LO. |
| Open Datasets | Yes | For KIP, CNP, DRP, we sample 1,000 instances according to the procedures specified in Tang et al. [54], Dragotto et al. [18], and Ghatkar et al. [27], respectively. |
| Dataset Splits | No | However, if the validation mean absolute error does not improve in 200 iterations, we terminate early. |
| Hardware Specification | Yes | The experiments for the benchmarks were run on a computing cluster with an Intel Xeon CPU E5-2683 and Nvidia Tesla P100 GPU with 64GB of RAM (for training). |
| Software Dependencies | Yes | Pytorch 2.0.1 [44] was used for all neural network models and scikit-learn 1.4.0 was used for gradient-boosted trees in the DNDP [46]. Gurobi 11.0.1 [30] was used as the MILP solver and gurobi-machinelearning 1.4.0 was used to embed the learning models into MILPs. |
| Experiment Setup | Yes | The sub-networks Ψd, Ψs, Ψv are feed-forward networks with one hidden layer of dimension 128. The decision-independent feature embedding dimension (m) is 64, and the instance embedding dimension (k) is 32. We use a batch size of 32, a learning rate of 0.01, and Adam [31] as an optimizer. |